OpenAI and MIT Media Lab Study on Affective Use of ChatGPT

OpenAI and the MIT Media Lab have released findings from a joint research effort to understand how "affective use"—interactions involving emotional engagement—impacts the social and emotional well-being of ChatGPT users. The research concludes that while emotional engagement with AI is rare for the majority of users, its effects on well-being are non-uniform and depend heavily on individual user behaviors, the modality of interaction (text vs. voice), and personal psychological factors.

Research Methodology

To capture both organic behavior and causal relationships, the researchers employed two parallel study designs:

  • Observational Study (OpenAI): An automated analysis of nearly 40 million ChatGPT interactions. To ensure user privacy, the pipeline used automated classifiers without human involvement to correlate self-reported user sentiment with conversation attributes.
  • Controlled Interventional Study (MIT Media Lab): An IRB-approved, pre-registered Randomized Controlled Trial (RCT) involving nearly 1,000 participants over four weeks. This study focused on the causal impact of platform features (personality and modality) and usage types on psychosocial states, including loneliness, emotional dependence, and problematic AI use.

Key Findings on Affective Use

Prevalence of Emotional Engagement

Emotional engagement with ChatGPT is rare in real-world usage. Affective cues—indicators of empathy, affection, or support—were absent in the vast majority of analyzed conversations. Among heavy users (defined as the top 1,000 daily users of Advanced Voice Mode), emotionally expressive interactions were concentrated in a small sub-population who were significantly more likely to view ChatGPT as a friend.

Impact of Modality: Voice vs. Text

Interaction modality produces mixed effects on emotional well-being:

  • Text: Users engaging via text exhibited more affective cues per message on average than voice users.
  • Voice: Voice modes were associated with improved well-being when used briefly, but led to worse outcomes with prolonged daily use.
  • Voice Personality: The use of a more "engaging" voice did not result in more negative outcomes compared to a neutral voice or text-based interaction.

Conversation Types and Well-being

The nature of the conversation significantly influences psychosocial outcomes:

  • Personal Conversations: These interactions, characterized by higher emotional expression from both user and model, were associated with higher levels of loneliness but lower emotional dependence and problematic use at moderate usage levels.
  • Non-Personal Conversations: These interactions tended to increase emotional dependence, particularly among heavy users.

User-Specific Risk Factors

While the study could not establish absolute causation for all personal factors, it identified strong correlations between negative outcomes and specific user traits:

  • Attachment Style: Individuals with a stronger tendency for attachment in relationships were more likely to experience negative effects.
  • Perception of AI: Users who viewed the AI as a friend that could fit into their personal life were more prone to negative outcomes.
  • Usage Duration: Extended daily use was consistently associated with worse emotional outcomes.

Study Limitations

The researchers noted several constraints that should be considered when interpreting these results:

  • Peer Review: The findings have not yet been peer-reviewed.
  • Scope: The study focused exclusively on English conversations with U.S. participants.
  • Data Sources: The research relied partly on self-reported survey data, which may not accurately reflect true experiences.
  • Technical Constraints: Automated classifiers used to identify affective cues are imperfect and may miss nuances.
  • Platform Specificity: The results are based on ChatGPT and may not generalize to other AI chatbot platforms.

Sources